iMAP: Discovering Complex Semantic Matches between Database Schemas

1
iMAP Discovering Complex Semantic Matches
between Database Schemas

• Robin Dhamankar, Yoonkyong Lee, AnHai Doan, Alon
  Halevy, Pedro Domingos
• Present by Trung Nguyen

2
Introduction

• Semantic mapping is crucial in combining
  disparate data sources
• Current works focus on 1-1 mappings
• This paper explore complex mappings
• Accomplished by
• Using searcher
• Exploiting domain knowledge

3
Motivating Example
4
Transferring Data from S to T

• Two steps
• Schema matching
• 1-1 mapping
• Complex mapping
• Create query expressions

5
iMAPs Architecture
6
Match Generator

• Discover a small set of candidate matches quickly
• Key idea using search
• Problem search space is infinite
• Solution
• Use different kind of searchers in parallel
• Special-purpose searchers can be added
• Use other techniques to prune out candidates

7
The Internal of a Searcher

• Search strategy
• How to search efficiently?
• ? Use beam search
• Match Evaluation
• How to evaluate match candidate?
• ? Apply different techniques machine learning,
  statistics, heuristics
• Termination Condition
• When to terminate?
• ? When the difference between Qk and Qk1 is less
  then some threshold

8
Types of Searchers (1)

• Text
• Numeric
• Category
• Schema-mismatch
• Unit conversion
• Date

9
Types of Searchers (2)
10
Similarity Estimator

• Further evaluates the candidates resulted from
  Match Generator
• Problem Accuracy reported by searchers may not
  be high
• Solution employs multiple evaluator modules
• Name-based evaluator
• Naïve Bayes evaluator

11
Match Selector

• Search for the best global match assignment
• Problem highest score candidate may not be the
  acceptable
• Solution
• Integrate with domain constraints

12
Exploiting Domain Knowledge

• To direct search process
• To prune out unqualified candidates

13
Types of Domain Knowledge

• Domain Constraints? Constraint implies on
• Single attribute in S
• Two attributes in S
• Multiple attributes in T
• Past Complex Matches
• Overlap Data
• External Data

14
Generating Explanations

• Be able to explain matches being made
• Help user gain insight into matching process
• User can provide feedback to converge the matches
  quickly

15
Type of User Questions

• Explain an existing matches
• Why match X is present in the output?
• Explain an absent match
• Why match Y is NOT present in the output?
• Explain match ranking
• Why match X is ranked higher than match Y in the
  output?

16
The Explanation Module

• Use dependency graph
• Dependency graph records the flow of matches,
  data and assumptions into and out of system
  components
• Nodes are
• Schema attributes
• Assumptions
• Candidate matches
• Domain knowledges pieces (e.g. constraints)
• Dependency graph in iMAP is small

17
A Sample Dependency Graph
18
Sample Explanations to User Questions
19
Empirical Evaluation (1)

• Domain and Data Sources
• Real Estate
• Inventory
• Cricket
• Financial Wizard
• Data Processing
• Experiments overlap and disjoint data

20
Empirical Evaluation (2)
21
Performance Result (1)

• Overall and 1- 1 matching accuracy

22
Performance Result (2)

• Complex matching accuracy Top-1

23
Performance Result (3)

• Complex matching accuracy Top-3

24
Performance Discussion

• Sometimes, iMAP adds noise to components ? reduce
  accuracy
• For disjoint data, discovering meaningful numeric
  relationship is very difficult
• Complex matches are not in top 1 but in top 3

25
Conclusion

• First system that explore complex mapping
• Employ searchers to help finding matches
• Utilize domain knowledge to guide the search and
  the evaluation process
• Offer a novel explanation facility

26
Thanks for your attention!Questions?